DeepSWE - Together.ai, in collaboration with Agentica, has open-sourced an AI Agent framework.

What is DeepSWE?

DeepSWE is an open-source AI agent framework developed by Together.ai in collaboration with Agentica. Based on the Qwen3-32B model, it is trained using reinforcement learning. DeepSWE performs exceptionally well in the SWE-Bench-Verified benchmark test, achieving an accuracy of 59.0% after Test-Time Extension (TTS) and a Pass@1 accuracy of 42.2% without TTS, ranking first among open-source agent frameworks. The model training data, code, training and evaluation logs are all open-source, helping developers learn and improve agents and promoting the application and development of reinforcement learning in software engineering.

DeepSWE's main functions

• Code understanding and editingTo understand and modify existing code to solve specific software problems or implement new features.
• Problem solvedBased on interaction with the environment, it solves complex software engineering problems, such as fixing issues on GitHub, implementing new code features, and debugging.
• Automated testing and verificationRun shell commands to build and test the code, verify the effectiveness of the solution, and ensure that the modified code does not break existing functionality.
• Multi-step reasoningWhen solving a task, we conduct multi-step reasoning and decision-making, gradually optimizing the solution until the task is finally completed.

DeepSWE's technical principles

• Reinforcement learning trainingIt is trained entirely from scratch using reinforcement learning (RL), without relying on distillation or supervised fine-tuning (SFT) of a stronger proprietary teacher model. Based on its interaction with the environment, it learns how to make optimal decisions in complex software engineering tasks.
• rLLM frameworkBased on the rLLM framework, this is a system used for later training of language agents, providing efficient data management and training processes, and supporting large-scale reinforcement learning training.
• Sparse result reward modelBased on the sparse result reward mechanism, a positive reward is only given when the code patch generated by the model passes all tests; otherwise, the reward is zero, which encourages the model to learn and generate high-quality solutions.
• Test-Time Extension (TTS)During the testing phase, multiple trajectories are generated, and the trajectory that correctly solves the problem is selected, significantly improving the model's accuracy. Combining the advantages of execution-based and non-execution-based validators achieves a substantial performance improvement.
• Kubernetes supportTo address the scaling challenges during training, Kubernetes support is integrated into the training environment to enable elastic scheduling and automatic scaling of containers, ensuring the efficiency and stability of the training process.

DeepSWE's project address

• HuggingFace model library:https://huggingface.co/agentica-org/DeepSWE-Preview

Application scenarios of DeepSWE

• Code understanding and editingDeepSWE can automatically understand and modify code, quickly locate and fix vulnerabilities, optimize performance bottlenecks, and refactor code structure, significantly improving the overall quality of the code.
• Software problem solvingEfficiently handle issues on GitHub, implement new features, and break down complex tasks into manageable subtasks, thereby quickly solving various problems in the field of software engineering.
• Automated testing and verificationDeepSWE can automatically generate test cases, execute code building and automated testing, perform regression testing, ensure the stability and reliability of software, and reduce the workload of manual testing.
• Multi-step reasoning and solution optimizationIt solves complex problems based on multi-step reasoning, continuously iterates and optimizes solutions, and accumulates knowledge and experience in the process of solving different problems, so as to better meet future challenges.
• Development assistance and collaborationDeepSWE provides developers with real-time code suggestions and completion hints, assists in project management and task allocation, promotes collaboration and communication among team members, and improves the efficiency of the entire development team.